Roger Lundqvist

Fast and robust extraction of hippocampus from MR images for diagnostics of Alzheimer's disease

Assessment of temporal lobe atrophy from magnetic resonance images is a part of clinical guidelines for the diagnosis of prodromal Alzheimers disease. As hippocampus is known to be among the first areas affected by the disease, fast and robust definition of hippocampus volume would be of great importance in the clinical decision making. We propose a method for computing automatically the volume of hippocampus using a modified multi-atlas segmentation framework, including an improved initialization of the framework and the correction of partial volume effect. The method produced a high similarity index, 0.87, and correlation coefficient, 0.94, with semi-automatically generated segmentations. When comparing hippocampus volumes extracted from 1.5T and 3T images, the absolute value of the difference was low: 3.2% of the volume. The correct classification rate for Alzheimers disease and cognitively normal cases was about 80% while the accuracy 65% was obtained for classifying stable and progressive mild cognitive impairment cases. The method was evaluated in three cohorts consisting altogether about 1000 cases, the main emphasis being in the analysis of the ADNI cohort. The computation time of the method is about 2 minutes on a standard laptop computer. The results show a clear potential for applying the method in clinical practice.

Implementation and Validation of an Adaptive Template Registration Method for 18F-Flutemetamol Imaging Data

The spatial normalization of PET amyloid imaging data is challenging because different white and gray matter patterns of negative (Aβ−) and positive (Aβ+) uptake could lead to systematic bias if a standard method is used. In this study, we propose the use of an adaptive template registration method to overcome this problem. Methods: Data from a phase II study (n = 72) were used to model amyloid deposition with the investigational PET imaging agent 18F-flutemetamol. Linear regression of voxel intensities on the standardized uptake value ratio (SUVR) in a neocortical composite region for all scans gave an intercept image and a slope image. We devised a method where an adaptive template image spanning the uptake range (the most Aβ− to the most Aβ+ image) can be generated through a linear combination of these 2 images and where the optimal template is selected as part of the registration process. We applied the method to the 18F-flutemetamol phase II data using a fixed volume of interest atlas to compute SUVRs. Validation was performed in several steps. The PET-only adaptive template registration method and the MR imaging–based method used in statistical parametric mapping were applied to spatially normalize PET and MR scans, respectively. Resulting transformations were applied to coregistered gray matter probability maps, and the quality of the registrations was assessed visually and quantitatively. For comparison of quantification results with an independent patient-space method, FreeSurfer was used to segment each subject’s MR scan and the parcellations were applied to the coregistered PET scans. We then correlated SUVRs for a composite neocortical region obtained with both methods. Furthermore, to investigate whether the 18F-flutemetamol model could be generalized to 11C-Pittsburgh compound B (11C-PIB), we applied the method to Australian Imaging, Biomarkers and Lifestyle (AIBL) 11C-PIB scans (n = 285) and compared the PET-only neocortical composite score with the corresponding score obtained with a semimanual method that made use of the subject’s MR images for the positioning of regions. Results: Spatial normalization was successful on all scans. Visual and quantitative comparison of the new PET-only method with the MR imaging–based method of statistical parametric mapping indicated that performance was similar in the cortical regions although the new PET-only method showed better registration in the cerebellum and pons reference region area. For the 18F-flutemetamol quantification, there was a strong correlation between the PET-only and FreeSurfer SUVRs (Pearson r = 0.96). We obtained a similar correlation for the AIBL 11C-PIB data (Pearson r = 0.94). Conclusion: The derived adaptive template registration method allows for robust, accurate, and fully automated quantification of uptake for 18F-flutemetamol and 11C-PIB scans without the use of MR imaging data.

Automated Quantification of 18F-Flutemetamol PET Activity for Categorizing Scans as Negative or Positive for Brain Amyloid: Concordance with Visual Image Reads

Clinical trials of the PET amyloid imaging agent 18F-flutemetamol have used visual assessment to classify PET scans as negative or positive for brain amyloid. However, quantification provides additional information about regional and global tracer uptake and may have utility for image assessment over time and across different centers. Using postmortem brain neuritic plaque density data as a truth standard to derive a standardized uptake value ratio (SUVR) threshold, we assessed a fully automated quantification method comparing visual and quantitative scan categorizations. We also compared the histopathology-derived SUVR threshold with one derived from healthy controls. Methods: Data from 345 consenting subjects enrolled in 8 prior clinical trials of 18F-flutemetamol injection were used. We grouped subjects into 3 cohorts: an autopsy cohort (n = 68) comprising terminally ill patients with postmortem confirmation of brain amyloid status; a test cohort (n = 172) comprising 33 patients with clinically probable Alzheimer disease, 80 patients with mild cognitive impairment, and 59 healthy volunteers; and a healthy cohort of 105 volunteers, used to define a reference range for SUVR. Visual image categorizations for comparison were from a previous study. A fully automated PET-only quantification method was used to compute regional neocortical SUVRs that were combined into a single composite SUVR. An SUVR threshold for classifying scans as positive or negative was derived by ranking the PET scans from the autopsy cohort based on their composite SUVR and comparing data with the standard of truth based on postmortem brain amyloid status for subjects in the autopsy cohort. The derived threshold was used to categorize the 172 scans in the test cohort as negative or positive, and results were compared with categorization using visual assessment. Different reference and composite region definitions were assessed. Threshold levels were also compared with corresponding thresholds derived from the healthy group. Results: Automated quantification (using pons as the reference region) demonstrated 91% sensitivity and 88% specificity and gave 3 false-positive and 4 false-negative scans. All 3 false-positive cases were either borderline-normal by standard of truth or had moderate to heavy cortical diffuse plaque burden. In the test cohort, the concordance between quantitative and visual read categorization ranged from 97.1% to 99.4% depending on the selection of reference and composite regions. The threshold derived from the healthy group was close to the histopathology-derived threshold. Conclusion: Categorization of 18F-flutemetamol amyloid imaging data using an automated PET-only quantification method showed good agreement with histopathologic classification of neuritic plaque density and a strong concordance with visual read results.

Combination of Biomarkers: PET [18F]Flutemetamol Imaging and Structural MRI in Dementia and Mild Cognitive Impairment

Background:The New National Institute on Aging-Alzheimer’s Association diagnostic guidelines for Alzheimer’s disease (AD) incorporate biomarkers in the diagnostic criteria and suggest division of biomarkers into two categories: Aβ accumulation and neuronal degeneration or injury. Objective:It was the aim of this study to compute hippocampus volume from MRI and a neocortical standard uptake value ratio (SUVR) from [18F]flutemetamol PET and investigate the performance of these biomarkers when used individually and when combined. Methods: Fully automated methods for hippocampus segmentation and for computation of neocortical SUVR were applied to MR and scans with the investigational imaging agent [18F]flutemetamol in a cohort comprising 27 AD patients, 25 healthy volunteers (HVs) and 20 subjects with amnestic mild cognitive impairment (MCI). Clinical follow-up was performed 2 years after the initial assessment. Results:Hippocampus volumes showed extensive overlap between AD and HV cases while PET SUVRs showed clear group clustering. When both measures were combined, there was a relatively compact cluster of HV scans and a less compact AD cluster. MCI cases had a bimodal distribution of SUVRs. [18F]Flutemetamol-positive MCI subjects showed a large variability in hippocampus volumes, indicating that these subjects were in different stages of neurodegeneration. Some [18F]flutemetamol-negative MCI scans had hippocampus volumes that were well below the HV range. Clinical follow-up showed that 8 of 9 MCI to AD converters came from the [18F]flutemetamol-positive group. Conclusion:Combining [18F]flutemetamol PET with structural MRI provides additional information for categorizing disease and potentially predicting shorter time to progression from MCI to AD, but this has to be validated in larger longitudinal studies.

Improved generation of probabilistic atlases for the expectation maximization classification

Probabilistic atlases present prior knowledge about the spatial distribution of various structures or tissues in a population, used commonly in segmentation. We propose three methods for generating probabilistic atlases: 1) the atlases are constructed in a template space using dense non-rigid transformations and transformed to the space of unseen data, 2) as the method 1 but atlas selection is performed in addition, and 3) the atlases are constructed directly in the space of the unseen data. The methods were evaluated in the segmentation of the hippocampus in 340 images from the Alzheimers Disease Neuroimaging Initiaitve (ADNI). Dice overlaps (similarity index, SI) were 0.84, 0.85 and 0.87 with reference segmentations and the correlation coefficients for the volumes were 0.84, 0.92 and 0.96 for the three methods tested. Our results show clearly the importance of probabilistic atlases in segmentation.

Hippocampal atrophy rate using an expectation maximization classifier with a disease-specific prior

Hippocampal atrophy is a well-known characteristic associated with Alzheimers disease. In this work, we propose a 4D Expectation Maximization framework for measuring the atrophy rate of the hippocampus from serial magnetic resonance images. One novelty of the framework is a disease-specific prior that regularizes the segmentation near the borders of the hippocampus. Regions where the hippocampus tends to get larger in the follow-up images than in the baseline are penalized. Using the ADNI cohort, we obtained classification accuracies of 83% for healthy control and Alzheimers disease patient groups and 60% for stable and progressive MCI groups using the baseline and 12-month follow-up images.

Fully automated quantification of [18F]flutemetamol images used to categorize scans into either normal or abnormal amyloid levels: Sensitivity and specificity against histopathology and concordance with visual read results

Constrained Laplacian-Based Automated Segmentation with Proximities algorithm from structural MRI and cortical thickness was defined using the tlink method. FDG uptake for each subject was registered to their structural MRI, interpolated onto their reconstructed cortical surfaces, normalized to average uptake of cerebellum and corrected partial volume effect by gray matter probability map. Feature selection The classification features were based on area under the receiver operating characteristic curve (AUC). We evaluated AUC of both of FDG uptake and cortical thickness in all nodes of cortical surface.We selected only node that demonstrated an AUC of 0.90 or better were retained. Classification using SVMFor classification, we used support vector machine (SVM) that find the maximum margin to optimally divide the NC and AD. The retained nodes were entered into SVM. Validation Classification performance was computed using leave-one-out crossvalidation. Results: In recent studies, each feature from FDG and cortical thickness demonstrated to inform diagnosis AD and NC. Therefore, combined features on this study will provide improvement of classification accuracy compared with results of previous single classifier. Conclusions: This study proposes a classification method for AD and NC based on combination of FDG-PETand cortical thickness. Our method will offer abilities using different data modalities. We believe that the different properties of multimodal images can provide understanding of AD pathology and more diagnostic accuracy than single classifier.